tray midtier 002 (#266)

Author: xTrayambak

docs/jit/AGNIPATH.md

@@ -0,0 +1,6 @@
+# Agnipath
+Agnipath (literally means "path of fire" in Hindi, I like to think of it as "hotpath") is the highest-tier JIT compiler in Bali. It aims to minimize boxing+unboxing and also produce the best quality code possible, no matter how much time it takes.
+
+It uses a SSA IR form to make certain optimizations trivial.
+
+**It currently does not exist, I am planning to work on it soon.**

shell.nix

@@ -13,6 +13,7 @@ mkShell {
     act
     boa
     mimalloc
+    ltrace
   ];
 
   LD_LIBRARY_PATH = lib.makeLibraryPath [

src/bali/internal/assembler/amd64.nim

@@ -194,6 +194,18 @@ const BaseAssemblerSize* {.intdefine: "BaliBaseAssemblerSize".} = 0x10000
 proc curAdr*(assembler: AssemblerX64): int64 =
   cast[int64](assembler.data) + assembler.offset
 
+proc dump*(s: AssemblerX64, file: string) =
+  assert(s.offset < s.capacity)
+
+  var data = newSeqOfCap[uint8](s.offset)
+  var i = 0
+
+  while i < s.offset:
+    data &= s.data[i]
+    inc i
+
+  writeFile(file, ensureMove(data))
+
 proc initAssemblerX64*(data: ptr UncheckedArray[byte], capacity: int): AssemblerX64 =
   result.data = data
   result.capacity = capacity
@@ -352,7 +364,7 @@ proc writeModrm[T, U](assembler: var AssemblerX64, rm: Rm[T], reg: U): int =
   of rmDirect:
     when T isnot void:
       assembler.writeField 0b11, byte(reg), byte(rm.directReg)
-      -1
+      return -1
     else:
       raiseAssert(
         "memory only operand. Direct register not allowed (how was this constructed?)"
@@ -387,7 +399,7 @@ proc writeModrm[T, U](assembler: var AssemblerX64, rm: Rm[T], reg: U): int =
       assembler.writeField 0b00, byte(reg), 0b100
       assembler.writeField byte(rm.simpleScale), byte(rm.simpleIndex), 0b101
       assembler.write rm.simpleDisp
-    -1
+    return -1
   of rmIndirectScaledAndBase:
     assert rm.baseIndex != regRsp, "rsp cannot be scaled"
     if rm.baseDisp == 0 and rm.base != regRbp and rm.base != regR13:
@@ -401,12 +413,12 @@ proc writeModrm[T, U](assembler: var AssemblerX64, rm: Rm[T], reg: U): int =
       assembler.writeField 0b10, byte(reg), 0b100
       assembler.writeField byte(rm.baseScale), byte(rm.baseIndex), byte(rm.base)
       assembler.write rm.baseDisp
-    -1
+    return -1
   of rmIndirectGlobal:
     assembler.writeField 0b00, byte(reg), 0b101
     let offset = assembler.offset
     assembler.write 0'i32
-    offset
+    return offset
 
 proc fixupRip[T](assembler: var AssemblerX64, modrm: Rm[T], location: int) =
   if location != -1:

src/bali/runtime/compiler/agnipath/types.nim

@@ -0,0 +1,3 @@
+## Types for Agnipath IR
+##
+## Copyright (C) 2025 Trayambak Rai (xtrayambak at disroot dot org)

src/bali/runtime/compiler/amd64/midtier.nim

@@ -323,10 +323,10 @@ proc compile*(
 
   var pipeline = Pipeline(fn: &lowered)
   pipeline.optimize(
-    {
+    @[
       Passes.NaiveDeadCodeElim, Passes.AlgebraicSimplification, Passes.EscapeAnalysis,
       Passes.CopyPropagation,
-    }
+    ]
   )
 
   return compileLowered(cgen, pipeline)

src/bali/runtime/compiler/madhyasthal/copy_propagation.nim

@@ -53,12 +53,16 @@ func eliminateUnmutatedCopies*(
         # Any attempts to eliminate this copy will cause semantic breakage.
         continue
 
-      if copy.dest notin pipeline.info.esc.locals:
+      if copy.source notin pipeline.info.esc.locals:
+        # debugecho "Can't optimize escapee %" & $copy.source & " -> %" & $copy.dest
         # If the register is not locally owned, we cannot
         # safely alias the copy as it might end up mutating
         # a global state, which'd cause semantic breakage.
         continue
 
+      # print pipeline.info.esc.locals
+      # debugecho "Optimizing local %" & $copy.source & " -> %" & $copy.dest
+
       unmutated.incl(copy)
 
   ensureMove(unmutated)

src/bali/runtime/compiler/madhyasthal/naive_dce.nim

@@ -9,6 +9,7 @@ func markDef(pipeline: var pipeline.Pipeline, reg: ir.Reg, at: SomeNumber) =
   pipeline.info.dce.defs.incl(
     Definition(reg: reg, inst: uint32(at))
   )
+  pipeline.info.dce.rawDefs.incl(reg)
   #!fmt: on
 
 func markUse(pipeline: var pipeline.Pipeline, reg: ir.Reg, at: SomeNumber) =
@@ -135,9 +136,19 @@ func scanAndElimDeadRefs*(pipeline: var pipeline.Pipeline, dead: HashSet[ir.Reg]
     pipeline.fn.insts &= inst
 
     if inst.args[0].kind == avkPos:
+      if inst.args[0].vreg notin pipeline.info.dce.rawDefs:
+        # We cannot determine if a global is alive,
+        # and we needn't care either.
+        continue
+
       pipeline.info.dce.alive.incl(inst.args[0].vreg)
 
     if inst.args[1].kind == avkPos:
+      if inst.args[1].vreg notin pipeline.info.dce.rawDefs:
+        # We cannot determine if a global is alive,
+        # and we needn't care either.
+        continue
+
       pipeline.info.dce.alive.incl(inst.args[1].vreg)
 
 func eliminateDeadCodeNaive*(pipeline: var pipeline.Pipeline) =

src/bali/runtime/compiler/madhyasthal/optimizer.nim

@@ -10,7 +10,7 @@ import
   ]
 import pkg/[shakar]
 
-func optimize*(pipeline: var pipeline.Pipeline, passes: set[Passes] = {}) =
+func optimize*(pipeline: var pipeline.Pipeline, passes: seq[Passes]) =
   for pass in passes:
     case pass
     of Passes.NaiveDeadCodeElim:

src/bali/runtime/compiler/madhyasthal/pipeline.nim

@@ -18,6 +18,7 @@ type
 
   DCEPassInfo* = object
     defs*: HashSet[Definition]
+    rawDefs*: HashSet[ir.Reg]
     uses*: HashSet[Use]
     alive*: HashSet[ir.Reg]
 
@@ -45,3 +46,8 @@ type
     EscapeAnalysis
 
   OptimizationPass* = proc(state: var Pipeline): bool
+
+func isUsedBeyond*(dce: DCEPassInfo, reg: ir.Reg, inst: uint32): bool {.inline.} =
+  for use in dce.uses:
+    if use.reg == reg and use.inst > inst:
+      return true

src/bali/runtime/vm/heap/bump_allocator.nim

@@ -36,4 +36,4 @@ proc release*(allocator: var BumpAllocator) =
 proc initBumpAllocator*(
     size: uint64 = DefaultAllocatorBufferSize
 ): BumpAllocator {.sideEffect.} =
-  result = BumpAllocator(pool: malloc(size), cap: size)
+  BumpAllocator(pool: malloc(size), cap: size)